Online Class-Incremental Continual Learning with Adversarial Shapley Value

1. A method for reducing a degree of catastrophic forgetting in a neural network by scoring training data samples according to an ability to preserve latent decision boundaries for previously observed classes while promoting learning from an input batch of new images from an online data stream, the method comprising: receiving the input batch of the new images from the online data stream; performing, with respect to a memory, a memory retrieval process comprising: obtaining an evaluation set for a first type of training data and a second type of training data from a first class-balanced random subset of the training data samples from the memory and a first candidate set from a second class-balanced random subset of the training data samples from the memory excluding any training data included in the second type of training data, wherein the evaluation set and the first candidate set comprise different data points, wherein the second type of training data corresponds to cooperative data points that are representative of training data samples in the memory to retain latent decision boundaries for previously observed classes, wherein the first type of training data corresponds to adversarial data points that are near samples in the input batch and with different labels to differentiate current classes from previously seen classes, and wherein the adversarial data points are adversarial to the new images from the online data stream; determining a K-Nearest Neighbor Shapley value (KNN-SV) of first candidate points among the first candidate set with respect to evaluation points among the evaluation set and the new images for the first type of training data and the second type of training data; selecting a subset of the first candidate points for memory replay to reduce the degree of catastrophic forgetting, by aggregating the determined KNN-SVs of the first candidate points, wherein a size of the subset of the first candidate points corresponds to a same size of the received input batch of the new images; concatenating the subset of the first candidate points selected for memory replay to the received input batch of new images to form a mini-batch for training the neural network with the formed mini-batch; and training the neural network to perform image recognition based on the formed mini-batch; and performing, with respect to the memory, a memory update process comprising: obtaining an evaluation set for a third type of training data from the first class-balanced random subset of the training data samples from the memory and a second candidate set from a randomly selected subset of the training data samples from the memory and the new images from the input batch, wherein a size of the second candidate set corresponds to a number of the new images in addition to a number of a size of the randomly selected training data samples from the memory; determining a KNN-SV of second candidate points among the second candidate set with respect to evaluation points among the evaluation set for the third type of training data by obtaining latent features of the third type of training data from the evaluation set and the second candidate set; determining a mean of the determined KNN-SVs of the second candidate points across the evaluation points; and replacing the second candidate points that are the training data samples in the memory having a smaller average KNN-SV than the training data samples from the input batch determined to have a higher average KNN-SV.

2. The method of claim 1, wherein the input batch corresponds to a set of new images sampled from the online data stream at a predefined time interval.

3. The method of claim 1, wherein the subset of the training data samples retrieved from memory uses the evaluation set and the first candidate set rather than the training data samples.

4. The method of claim 1, wherein the K-Nearest Neighbor Shapley value (KNN-SV) of first candidate points among the first candidate set with respect to the evaluation points among the evaluation set and the new images for the first type of training data and the second type of training data are determined by obtaining latent features of the evaluation set and the first candidate set, wherein a sign and a magnitude of the KNN-SV indicates an equivalence in class labels and a relative similarity of a particular candidate point and an particular evaluation point.

5. The method of claim 1, wherein the first class-balanced random subset of the training data samples is balanced in accordance with a number of examples from each class present in the memory.

6. The method of claim 1, wherein a size of the evaluation set for the first type of training data corresponds to a size of a number of the new images, wherein a size of the evaluation set for the second type of training data and a size of the first candidate set corresponds to a number of samples per class.

7. The method of claim 1, wherein the determined KNN-SVs of the first candidate points is aggregated according to an adversarial Shapley Value (ASV) or a mean-variation (ASVμ), wherein the ASV corresponds to a single maximum value with respect to Type 2 evaluation points minus a single minimum value with respect to Type 1 evaluation points and the ASVμ corresponds to an average value with respect to the Type 2 evaluation points minus an average value with respect to the Type 1 evaluation points.

8. The method of claim 1, wherein the first type of training data corresponds to a first group of data samples among the subset of the training data samples with a negative average KNN-SV of a large magnitude with respect to Type 1 evaluation set and the second type of training data corresponds to a second group of data samples among the subset of the training data samples with a positive average KNN-SV of a large magnitude with respect to Type 2 evaluation set.

9. The method of claim 1, wherein the randomly selected subset of the training data samples from the memory along with the new images is set based on a uniform random sampling.

10. A neural network system comprising: one or more processors; and a non-transitory memory storing instructions for reducing a degree of catastrophic forgetting in a neural network by scoring training data samples according to an ability to preserve latent decision boundaries for previously observed classes while promoting learning from an input batch of new images from an online data stream, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform: receiving the input batch of the new images from the online data stream; obtaining an evaluation set for a first type of training data and a second type of training data from a first class-balanced random subset of the training data samples from the memory and a first candidate set from a second class-balanced random subset of the training data samples from the memory excluding any training data included in the second type of training data, wherein the evaluation set and the first candidate set comprise different data points, wherein the second type of training data corresponds to cooperative data points that are representative of training data samples in the memory to retain latent decision boundaries for previously observed classes, wherein the first type of training data corresponds to adversarial data points that are near samples in the input batch and with different labels to differentiate current classes from previously seen classes, and wherein the adversarial data points are adversarial to the new images from the online data stream; determining a K-Nearest Neighbor Shapley value (KNN-SV) of first candidate points among the first candidate set with respect to evaluation points among the evaluation set and the new images for the first type of training data and the second type of training data; selecting a subset of the first candidate points for memory replay to reduce the degree of catastrophic forgetting, by aggregating the determined KNN-SVs of the first candidate points, wherein a size of the subset of the first candidate points corresponds to a same size of the received input batch of the new images; concatenating the subset of the first candidate points selected for memory replay to the received input batch of new images to form a mini-batch for training the neural network with the formed mini-batch; and training the neural network to perform image recognition based on the formed mini-batch; and a memory update process comprising: obtaining an evaluation set for a third type of training data from the first class-balanced random subset of the training data samples from the memory and a second candidate set from a randomly selected subset of the training data samples from the memory and the new images from the input batch, wherein a size of the second candidate set corresponds to a number of the new images in addition to a number of a size of the randomly selected training data samples from the memory; determining a KNN-SV of second candidate points among the second candidate set with respect to the evaluation points by obtaining latent features of the third type of training data from the evaluation set and the second candidate set; determining a mean of the determined KNN-SVs of the second candidate points across the evaluation points; and replacing the second candidate points that are the training data samples in the memory having a smaller average KNN-SV than the training data samples from the input batch determined to have a higher average KNN-SV.

11. The neural network system of claim 10, wherein the input batch corresponds to a set of new images sampled from the online data stream at a predefined time interval.

12. The neural network system of claim 10, wherein the subset of the training data samples retrieved from memory uses the evaluation set and the first candidate set rather than the training data samples.

13. The neural network system of claim 10, wherein the K-Nearest Neighbor Shapley value (KNN-SV) of first candidate points among the first candidate set with respect to evaluation points among the evaluation set and the new images for the first type of training data and the second type of training data are determined by obtaining latent features of the evaluation set and the first candidate set, wherein a sign and a magnitude of the KNN-SV indicates an equivalence in class labels and a relative similarity of a particular candidate point and an particular evaluation point.

14. The neural network system of claim 10, wherein the first class-balanced random subset of the training data samples is balanced in accordance with a number of examples from each class present in the memory.

15. A non-transitory memory storing one or more programs, which, when executed by one or more processors of a device, cause the device to be configured to perform: receiving an input batch of new images from an online data stream; performing a memory retrieval process comprising: obtaining an evaluation set for a first type of training data and a second type of training data from a first class-balanced random subset of training data samples from the memory and a first candidate set from a second class-balanced random subset of the training data samples from the memory excluding any training data included in the second type of training data, wherein the evaluation set and the first candidate set comprise different data points, wherein the second type of training data corresponds to cooperative data points that are representative of training data samples in the memory to retain latent decision boundaries for previously observed classes, wherein the first type of training data corresponds to adversarial data points that are near samples in the input batch and with different labels to differentiate current classes from previously seen classes, and wherein the adversarial data points are adversarial to the new images from the online data stream; determining a K-Nearest Neighbor Shapley value (KNN-SV) of first candidate points among the first candidate set with respect to evaluation points among the evaluation set and the new images for the first type of training data and the second type of training data; selecting a subset of the first candidate points for memory replay to reduce a degree of catastrophic forgetting in a neural network, by aggregating the determined KNN-SVs of the first candidate points, wherein a size of the subset of the first candidate points corresponds to a same size of the received input batch of the new images; concatenating the subset of the first candidate points selected for memory replay to the received input batch of new images to form a mini-batch for training the neural network with the formed mini-batch; and training the neural network to perform image recognition based on the formed mini-batch; and performing a memory update process comprising: obtaining an evaluation set for a third type of training data from the first class-balanced random subset of the training data samples from the memory and a second candidate set from a randomly selected subset of the training data samples from the memory and the new images from the input batch, wherein a size of the second candidate set corresponds to a number of the new images in addition to a number of a size of the randomly selected training data samples from the memory; determining a KNN-SV of second candidate points among the second candidate set with respect to the evaluation points by obtaining latent features of the third type of training data from the evaluation set and the second candidate set; determining a mean of the determined KNN-SVs of the second candidate points across the evaluation points; and replacing the second candidate points that are the training data samples in the memory having a smaller average KNN-SV than the training data samples from the input batch determined to have a higher average KNN-SV.

16. The non-transitory memory of claim 15, wherein the input batch corresponds to a set of new images sampled from the online data stream at a predefined time interval.

17. The non-transitory memory of claim 15, wherein the subset of the training data samples retrieved from memory uses the evaluation set and the first candidate set rather than the training data samples.

18. The non-transitory memory of claim 15, wherein the K-Nearest Neighbor Shapley value (KNN-SV) of first candidate points among the first candidate set with respect to evaluation points among the evaluation set and the new images for the first type of training data and the second type of training data are determined by obtaining latent features of the evaluation set and the first candidate set, wherein a sign and a magnitude of the KNN-SV indicates an equivalence in class labels and a relative similarity of a particular candidate point and an particular evaluation point.

19. The non-transitory memory of claim 15, wherein the first class-balanced random subset of the training data samples is balanced in accordance with a number of examples from each class present in the memory.